Study On Target Recognition In SAR Image Via The Monogenic Signal

Synthetic aperture radar(SAR)is an active sensor worked in the microwave band.It could work in 24-hour a day and all-weather,due to the ability not to be limited by the lighting and environment conditions.With the development of ultra-wideband imaging technique,SAR could penetrate the earth surface,vegetation and hence get the valuable information.Automatic target recognition(ATR)is a basic research topic for SAR image interpretation.It plays an important role in battlefield surveillance,anti-ballistic missile,air defense,strategic warning,etc,and hence achieves the perceive of battlefield situation and the generation of important intelligence.Though a great many works have been done previously,SAR ATR is still an open problem,especially for various extended operating conditions.As widely reported,target scattering phenomenology is susceptible to the variation of target pose,radar depression angle,configuration,articulation and occlusion,noise corruption,and hence results in the mutable characteristics of SAR image.It is therefore necessary to make full consideration on these issues.By deeply analyzing the preceding works on SAR target recognition,this dissertation comes the conclusion that the crux of target recognition lies in two procedures,the representation of target scattering characteristics and the implementation of classification learning.To improve recognition performance,this dissertation introduces two recently developed techniques,the monogenic signal and sparse signal modeling.Specifically,the monogenic signal is developed to achieve the representation of target scattering phenomenology,while sparse signal modeling is modified to perform efficient classification.It is expected to promote target recognition performance by advanced learning skill and effective feature descriptor.To jointly consider the components of multi-resolution monogenic signal into the framework of sparse signal modeling,this dissertation presents three different frameworks,including information fusion in the spatial domain,composite kernel learning in Hilbert space,and Riemannian manifold learning.The contribution and novelty of this dissertation include:· First,this dissertation comprehensively reviews the development of SAR ATR,especially for the famous three-stage processing framework.The current approaches to SAR ATR are categorized as template matching scheme,model-based vision,and the improved strategy.Both advantages and disadvantages of each algorithms have been summarized,followed by the generalization of the crucial issues.To solve these issues,the dissertation introduces two recently developed techniques,sparse signal modeling and the monogenic signal.The former is used to improve the accuracy of classification learning,while the latter is employed to capture the characteristics of target scattering phenomenology.· Second,this dissertation provides the development of sparse signal modeling,as well as compressed sensing.The related background knowledges include the work mechanism,the prerequisites,and the reconstruction algorithms.Afterward,this dissertation gives the applicability of sparse signal modeling in SAR image interpretation,including image super-resolution,image recovery,speckle reduction,image compression,terrain classification.To adopt sparse signal modeling for target recognition in SAR image,this dissertation proposes two modified schemes,i.e.,performing sparse representation via Fourier spectrum descriptor and reaching the inference via Dempster-Shafer's theory of evidence.Improvement of signal representation via Fourier transform.To deal with signal alignment and noise corruption,this dissertation proposes to implement sparse representation by Fourier spectrum bands of signal.Since signal energy in the frequency domain concentrates in a small portion of frequency bands,they are discriminative and hence used to represent the original signal.The proposed scheme is then used to feed into the framework of sparse signal modeling.The decision is made according to the characteristics of representation on reconstruction.Improvement of decision rule via Dempster-Shafer's theory of evidence.To improve the reliability of decision rule,this dissertation recommends a soft decision via Dempster-Shafer's theory of evidence.The imprecision on uncertainty measurement is modeled by sample-wise ambiguity and the class-wise ambiguity during the quantification of probability mass.Various pieces of evidence derived from the candidate samples are pooled by means of Dempster's rule of combination,from which an inference can be reached.· Third,to cope with target recognition under the extended operating conditions,the monogenic signal is proposed.Log-Gabor filter bank is employed to implement multi-resolution monogenic signal.To improve recognition accuracy,this dissertation jointly considers the components derived from multi-resolution monogenic signal into the framework of sparse signal modeling,by which three different strategies are developed.The 1stStrategy—Spatial Information Aggregation To aggregate various information derived from the monogenic signal,this dissertation designs a novel framework by performing information fusion in the spatial domain.According to the presented thought,three implementation schemes,feature-level fusion,score-level fusion,and multi-feature & multi-task joint sparse representation are recommended.The 2ndStrategy—Composite Kernel Learning To deal with the dataset whose target classes are not linearly separable,this dissertation proposes to achieve classification by composite kernel learning.According to this thought,two implementation algorithms,stacked features and summation kernels are developed.These algorithms reflect the thoughts of mapping followed by feature combination,and kernel combination followed by mapping.The 3rdStrategy—Riemannian Manifold Learning Recent studies show that great advantages can be achieved by considering the batch of signals with nonEuclidean geometry.Therefore,this dissertation proposes to perform target recognition with manifold learning skill.By apply the monogenic signal into the framework of manifold learning,thee strategies,Riemanian manifold learning on symmetry positive definite matrix space,Grassmann manifold learning via the components of multi-resolution monogenic signal,and Grassmann manifold learning via steerable Riesz wavelet frames are developed.· Fourth,multiple comparative experiments have been performed on MSTAR SAR database to validate the proposed algorithms.